Additionally, GPU-based images also have the following requirements:
BatchX offers GPU support to the images via the NVIDIA Container Toolkit. Images are required to include and make use of the CUDA toolkit in order to access underlying GPU capabilities .
The easiest way to ensure you have a GPU-supported image is to use
nvidia/cuda as base image.
If for example, you want to run a machine learning workflow to train a neural network on Tensorflow 2.1.0 framework, then you could use as base image
tensorflow/tensorflow:2.1.0-gpu,which is based on a Linux base image including the CUDA toolkit .
GPU-based images must declare a
runtime/gpus property in their manifest. This property can take the following values:
supported: Jobs might use GPUs, but it's not mandatory.
required: Jobs at least need one GPU to be executed.
If this manifest field is omitted, BatchX won't allow running this image with a
-g parameter set.